I’m very excited to share our new work titled “NeRN - Learning Neural Representations for Neural Networks”. Preprint and code are available now😊🧵
@ZoharRimon@ronvain@levishir667@EladRichardson@PinkyMintz
Paper: https://t.co/TUYgOhVrxk
Code: https://t.co/dYHqyHIY7h
[1/9]
Animation 🤝 Robotics
ProtoMotions GTC 2026 release — bridging the gap between digital humans and real humanoid robots.
Train in simulation. Deploy on hardware. One framework, one codebase.
https://t.co/33aV9hj4DH
Just arrived at San Diego for #Neurips2025 ✈️
If you're interested in
- Tactile sensing 🫰
- Uncertain decision making ↔️
- Medical robotics 🩺
Reach out or come see if you process touch better than AI at our poster on Fri 16:30-19:30 #2206.
@NeurIPSConf
1. Train a behavior foundation model with a rich token-based representation (eg, MaskedMimic).
2. Fine tune a state-based tokenizer using RL, leveraging the BFM for meaningful gradients.
=> Converge 6x faster, 50x less params, dramatically improved robustness.
Behavior models are great — but getting them to do *exactly* what you want takes *tons* of prompt hacking (just like vision-language models) 😪
No more, with our new work 📰 -
Task Tokens: A Flexible Approach to Adapting Behavior Foundation Models
✨The potential of Task Tokens is truly exciting! Imagine seamlessly adapting behavior models to various tasks with just a few learned params - 5 tasks using just ~1M vs standard HRL using ~125M.
This is a promising step towards more flexible and generalizable AI behavior😊
10/10
🚀Using BFMs to learn a new task with less than 0.1% additional params while still enabling user input??🚀
Task Tokens: A Flexible Approach to Adapting Behavior Foundation Models
📄: https://t.co/Q5sA9eMhUM
💻: https://t.co/N8TDfJDms5
w/@ZoharRimon@shiemannor@ChenTessler
🧵1/10
As our base BFM we use the great work MaskedMimic by Tessler et al. https://t.co/uJMAege9Wz, although our approach can be extended to other BFMs with a transformer-based architecture
9/10
Parallels between AI and #LK99:
- Both are modern-day alchemy. Just try more recipes until eureka.
- The holy grail is simpler than we expected.
- Lots of hyperparameters to tune.
- Random seeds matter.
- Arxiv is the new battlefield.
- Scaling up is the key.
- Tend to break the internet with every update.
- Fastest growth in Twitter experts.
- Democratization: there will be GPT-4's from big corps, and Alpaca's from grassroots.
לייבטויט מהשיחה של סאם אלטמן באונ' תל אביב.
לא בלייב, כי לא הייתי שם.
רואה את ההקלטה וזורק כאן את כל המחשבות שלי פלוס תמלול.
אז:
סם אלטמן, מנכ"ל OpenAI והשם שחלקכם פחות מכירים - איליה סוצקבר, המדען הראשי של OpenAI שהוא ישראלי לשעבר שלמד פה בארץ. >>
https://t.co/JgJnLWBRGK
"Pyrallis, a year has passed
And your library's made coding a blast
Configuring projects is now so swift
And it's all thanks to your magical gift"
- ChatGPT
Pyrallis is celebrating its first birthday with 20K+ installs 😊🎂
Join the configvolution 🐉https://t.co/XTJcZKBna5